Fuller S.H., Millett L.I. The Future of Computing Performance: Game Over or Next Level?

Подождите немного. Документ загружается.

The Future of Computing Performance: Game Over or Next Level?

46 THE FUTURE OF COMPUTING PERFORMANCE

· Real-time decision-making—enables growing use of computa-

tional assistance for various complex problem-solving tasks, such

as speech transcription and language translation.

· Collaboration technology—offers a more immersive and interac-

tive 3D environment for real-time collaboration and telepresence.

· Machine-learning algorithms—ﬁlter e-mail spam, supply reli-

able telephone-answering services, and make book and music

recommendations.

Computers have become so pervasive that a vast majority of end-

users are not computer aﬁcionados or system experts; rather, they are

experts in some other ﬁeld or disciplines, such as science, art, education,

or entertainment. The shift has challenged the efﬁciency of human-com-

puter interfaces. There has always been an inherent gap between a user’s

conceptual model of a problem and a computer’s model of the problem.

However, given the change in demographics of computer users, the need

to bridge the gap is now more acute than ever before. The increased

complexity of common end-user tasks (“ﬁnd a picture like this” rather

than “add these two numbers”) and the growing need to be able to offer

an effective interface to a non-computer-expert user at a higher level of

object semantics (for example, presenting not a Fourier transform data

dump of a ﬂower image but a synthesized realistic visual of a ﬂower)

have together increased the computational capability needed to provide

real-time responses to user actions.

Bridging the gap would be well served by computers that can deal

with natural user inputs, such as speech and gestures, and output content

in a visually rich form close to that of the physical world around us. A typ-

ical everyday problem requires multiple iterations of execute and evaluate

between the user and the computer system. Each such iteration normally

narrows the original modeling gap, and this in turn requires additional

computational capability. The larger the original gap, the more computa-

tion is needed to bridge it. For example, some technology-savvy users

working on an image-editing problem may iterate by editing a low-level

machine representation of an image, whereas a more typical end-user may

interact only at the level of a photo-real output of the image with virtual

brushes and paints.

Thanks to sustained growth in computing performance over the

years, more effective computer-use models and visually rich human-

computer interfaces are introducing new potential ways to bridge the gap.

An alternative to involving the end-user in each iteration is to depend

on a computer’s ability to reﬁne model instances by itself and to nest

multiple iterations of such an analytics loop for each iteration of a visual

computing loop involving an end-user. Such nesting allows a reduction in

The Future of Computing Performance: Game Over or Next Level?

THE NEED FOR CONTINUED PERFORMANCE GROWTH 47

the number of interactions between a user and his or her computer and

therefore an increase in the system’s efﬁciency or response. However, it

also creates the need to sustain continued growth in computational per-

formance so that a wider variety of more complex tasks can be simulated

and solved in real time for the growing majority of end-users. Real-time

physical and behavioral simulation of even a simple daily-life object or

events (such as water ﬂow, the trajectory of a ball in a game, and sum-

marizing of a text) is a surprisingly computationally expensive task, and

requires multiple iterations or solutions of a large number of subproblems

derived from decomposition of the original problem.

Computationally intensive consumer applications include such phe-

nomena as virtual world simulations and immersive social-networking,

video karaoke (and other sorts of real-time video interactions), remote

education and training that require simulation, and telemedicine (includ-

ing interventional medical imaging).

THE IMPORTANCE OF COMPUTING PERFORMANCE

FOR ENTERPRISE PRODUCTIVITY

Advances in computing technology in the form of more convenient

communication and sharing of information have favorably affected the

productivity of enterprises. Improved communication and sharing have

been hallmarks of computing from the earliest days of time-sharing in

corporate or academic environments to today’s increasingly mobile, smart

phone-addicted labor force. Younger employees in many companies today

can hardly recall business processes that did not make use of e-mail, chat

and text messaging, group calendars, internal Web resources, blogs, Wiki

toolkits, audio and video conferencing, and automated management of

workﬂow. At the same time, huge improvements in magnetic storage

technology, particularly for disk drives, have made it affordable to keep

every item of an organization’s information accessible on line. Individual

worker productivity is not the only aspect of an enterprise that has been

affected by continued growth in computing performance. The ability of

virtually every sort of enterprise to use computation to understand data

related to its core lines of business—sometimes referred to as analytics—

has improved dramatically as computer performance has increased over

the years. In addition, massive amounts of data and computational capa-

bility accessible on the Internet have increased the demand for Web ser-

vices, or “software as a service,” in a variety of sectors. Analytics and

For more on emerging applications and their need for computational capability, see

Justin Rattner, 2009, The dawn of terascale computing, IEEE Solid-State Circuits Magazine

1(1): 83-89.

The Future of Computing Performance: Game Over or Next Level?

48 THE FUTURE OF COMPUTING PERFORMANCE

the implications of Web services for computing performance needs are

discussed below.

Analytics

Increases in computing capability and efﬁciency have made it feasible

to perform deep analysis of numerous kinds of business data—not just

off line but increasingly in real time—to obtain better input into business

decisions.

Efﬁcient computerized interactions between organizations

have created more efﬁcient end-to-end manufacturing processes through

the use of supply-chain management systems that optimize invento-

ries, expedite product delivery, and reduce exposure to varying market

conditions.

In the past, real-time business performance needs were dictated

mostly by transaction rates. Analytics (which can be thought of as com-

putationally enhanced decision-making) were mostly off line. The com-

putational cost of actionable data-mining was too high to be of any value

under real-time use constraints. However, the growth in computing per-

formance has now made real-time analytics affordable for a larger class

of enterprise users.

One example is medical-imaging analytics. Over the last 2 decades,

unprecedented growth has taken place in the amount and complexity of

digital medical-image data collected on patients in standard medical prac-

tice. The clinical necessity to diagnose diseases accurately and develop

treatment strategies in a minimally invasive manner has mandated the

development of new image-acquisition methods, high-resolution acquisi-

tion hardware, and novel imaging modalities. Those requirements have

placed substantial computational burdens on the ability to use the image

information synergistically. With the increase in the quality and utility of

medical-image data, clinicians are under increasing pressure to generate

more accurate diagnoses or therapy plans. To meet the needs of the clini-

cian, the imaging-research community must provide real-time (or near

real-time) high-volume visualization and analyses of the image data to

optimize the clinical experience. Today, nearly all use of computation in

medical imaging is limited to “diagnostic imaging.” However, with sufﬁ-

cient computational capability, it is likely that real-time medical interven-

tions could become possible. The shift from diagnostic imaging to inter-

ventional imaging can usher in a new era in medical imaging. Real-time

IBM’s Smart Analytics System, for example, is developing solutions aimed at retail,

insurance, banking, health care, and telecommunication. For more information see the IBM

Smart Analytics System website, available online at http://www-01.ibm.com/software/

data/infosphere/smart-analytics-system/.

The Future of Computing Performance: Game Over or Next Level?

THE NEED FOR CONTINUED PERFORMANCE GROWTH 49

medical analytics can guide medical professionals, such as surgeons, in

their tasks. For example, surface extractions from volumetric data coupled

with simulations of various what-if scenarios accomplished in real time

offer clear advantages over basic preoperative planning scenarios.

Web Services

In the last 15 years, the Internet and the Web have had a transforma-

tional effect on people’s lives. That effect has been enabled by two concur-

rent and interdependent phenomena: the rapid expansion of Internet con-

nectivity, particularly high-speed Internet connections, and the emergence

of several extraordinarily useful Internet-based services. Web search and

free Web-based e-mail were among the ﬁrst such services to explode

in popularity, and their emergence and continuous improvements have

been made possible by dramatic advances in computing performance,

storage, and networking technologies. Well beyond text, Web-server data

now include videos, photos, and various other kinds of media. Users—

individuals and businesses—increasingly need information systems to

see data the way they do, identify what is useful, and assemble it for

them. The ability to have computers understand the data and help us

to use it in various enterprise endeavors could have enormous beneﬁts.

As a result, the Web is shifting its focus from data presentation to end-

users to automatic data-processing on behalf of end-users. Finding pre-

ferred travel routes while taking real-time trafﬁc feeds into account and

rapid growth in program trading are some of the examples of real-time

decision-making.

Consider Web search as an example. A Web search service’s funda-

mental task is to take a user’s query, traverse data structures that are effec-

tively proportional in size to the total amount of information available on

line, and decide how to select from among possibly millions of candidate

results the handful that would be most likely to match the user’s expecta-

tion. The task needs to be accomplished in a few hundred milliseconds in

a system that can sustain a throughput of several thousand requests per

second. This and many other Web services are offered free and rely on

on-line advertisement revenues, which, depending on the service, may

bring only a few dollars for every thousand user page views. The com-

puting system that can meet those performance requirements needs to be

not only extremely powerful but also extremely cost-efﬁcient so that the

business model behind the Internet service remains viable.

The appetite of Internet services for additional computing perfor-

mance doesn’t appear to have a foreseeable limit. A Web search can be

used to illustrate that, although a similar rationale could be applied to

other types of services. Search-computing demands fundamentally grow

The Future of Computing Performance: Game Over or Next Level?

50 THE FUTURE OF COMPUTING PERFORMANCE

in three dimensions: data-repository increases, search-query increases,

and service-quality improvements. The amount of information currently

indexed by search engines, although massive, is still generally considered

a fraction of all on-line content even while the Web itself keeps expanding.

Moreover, there are still several non-Web data sources that have yet to

be added to the typical Web-search repositories (such as printed media).

Universal search,

for example, is one way in which search-computing

demands can dramatically increase as all search queries are simultane-

ously sent to diverse data sources. As more users go online or become

more continuously connected to the Internet through better wireless links,

trafﬁc to useful services would undergo further substantial increases.

In addition to the amount of data and types of queries, increases

in the quality of the search product invariably cause more work to be

performed on behalf of each query. For example, better results for a

user’s query will often be satisﬁed by searching also for some common

synonyms or plurals of the original query terms entered. To achieve the

better results, one will need to perform multiple repository lookups for

the combinations of variations and pick the best results among them, a

process that can easily increase the computing demands for each query

by substantial factors.

In some cases, substantial service-quality improvements will demand

improvements in computing performance along multiple dimensions

simultaneously. For example, the Web would be much more useful if

there were no language barriers; all information should be available in

every existing language, and this might be achievable through machine-

translation technology at a substantial processing cost. The cost would

come both from the translation step itself, because accurate translations

require very large models or learning over large corpora, and from the

increased amount of information that then becomes available for users of

every language. For example, a user search in Italian would traverse not

only Italian-language documents but potentially documents in every lan-

guage available to the translation system. The beneﬁts to society at large

from overcoming language barriers would arguably rival any other single

technologic achievement in human history, especially if they extended to

speech-to-speech real-time systems.

The prospect of mobile computing systems—such as cell phones,

vehicle computers, and media players—that are increasingly powerful,

ubiquitous, and interconnected adds another set of opportunities for bet-

See Google’s announcement: Google begins move to universal search: Google introduces

new search features and unveils new homepage design,” Press Release, Google.com, May

16, 2007, available online at http://www.google.com/intl/en/press/pressrel/universal-

search_20070516.html.

The Future of Computing Performance: Game Over or Next Level?

THE NEED FOR CONTINUED PERFORMANCE GROWTH 51

ter computing services that go beyond simply accessing the Web on more

devices. Such devices could act as useful sensors and provide a rich set

of data about their environment that could be useful once aggregated for

real-time disaster response, trafﬁc-congestion relief, and as-yet-unimag-

ined applications. An early example of the potential use of such systems

is illustrated in a recent experiment conducted by the University of Cali-

fornia, Berkeley, and Nokia in which cell phones equipped with GPS units

were used to provide data for a highway-conditions service.

More generally, the unabated growth in digital data, although still

a challenge for managing and sifting, has now reached a data volume

large enough in many cases to have radical computing implications.

Such huge amounts of data will be especially useful for a class of prob-

lems that have so far deﬁed analytic formulation and been reliant on a

statistical data-driven approach. In the past, because of insufﬁciently large

datasets, the problems have had to rely on various, sometimes question-

able heuristics. Now, the digital-data volume for many of the problems

has reached a level sufﬁcient to revert to statistical approaches. Using sta-

tistical approaches for this class of problems presents an unprecedented

opportunity in the history of computing: the intersection of massive data

with massive computational capability.

In addition to the possibility of solving problems that have heretofore

been intractable, the massive amounts of data that are increasingly avail-

able for analysis by small and large businesses offer the opportunity to

develop new products and services based on that analysis. Services can

be envisioned that automate the analysis itself so that the businesses do

not have to climb this learning curve. The machine-learning community

has many ideas for quasi-intelligent automated agents that can roam the

Web and assemble a much more thorough status of any topic at a much

deeper level than a human has time or patience to acquire. Automated

inferences can be drawn that show connections that have heretofore been

unearthed only by very talented and experienced humans.

On top of the massive amounts of data being created daily and all

that portends for computational needs, the combination of three elements

has the potential to deliver a massive increase in real-time computa-

tional resources targeted toward end-user devices constrained by cost

and power:

See the University of California, Berkeley, press release about this experiment (Sarah

Yang, 2008, Joint Nokia research project captures trafﬁc data using GPS-enabled cell phones,

Press Release, UC Berkeley News, February 8, 2008, available online at http://berkeley.edu/

news/media/releases/2008/02/08_gps.shtml).

Wired.com ran a piece in 2008 declaring “the end of science”: The Petabyte Age: Be-

cause more isn’t just more—more is different,” Wired.com, June 23, 2008, available online

at http://www.wired.com/wired/issue/16-07.

The Future of Computing Performance: Game Over or Next Level?

52 THE FUTURE OF COMPUTING PERFORMANCE

· Clouds of servers.

· Vastly larger numbers of end-user devices, consoles, and various

form-factor computing platforms.

· The ubiquitous connectivity of computing equipment over a ser-

vice-oriented infrastructure backbone.

The primary technical challenge to take advantage of those resources

lies in software. Speciﬁcally, innovation is needed to enable the discovery

of the computing needs of various functional components of a speciﬁc ser-

vice offering. Such discovery is best done adaptively and under the real-

time constraints of available computing bandwidth at the client-server

ends, network bandwidth, and latency. On-line games, such as Second Life,

and virtual world simulations, such as Google Earth, are examples of such

a service. The services involve judicious decomposition of computing

needs over public client-server networks to produce an interactive, visu-

ally rich end-user experience. The realization of such a vision of connected

computing will require not only increased computing performance but

standardization of network software layers. Standardization should make

it easy to build and share unstructured data and application program-

ming interfaces (APIs) and enable ad hoc and innovative combinations

of various service offerings.

In summary, computing in a typical end-user’s life is undergoing a

momentous transformation from being useful yet nonessential software

and products to being the foundation for around-the-clock relied-on vital

services delivered by tomorrow’s enterprises.

The Future of Computing Performance: Game Over or Next Level?

What Is Computer Performance?

ast, inexpensive computers are now essential to numerous human

endeavors. But less well understood is the need not just for fast

computers but also for ever-faster and higher-performing comput-

ers at the same or better costs. Exponential growth of the type and scale

that have fueled the entire information technology industry is ending.

In addition, a growing performance gap between processor performance

and memory bandwidth, thermal-power challenges and increasingly

expensive energy use, threats to the historical rate of increase in transistor

density, and a broad new class of computing applications pose a wide-

ranging new set of challenges to the computer industry. Meanwhile, soci-

etal expectations for increased technology performance continue apace

and show no signs of slowing, and this underscores the need for ways

to sustain exponentially increasing performance in multiple dimensions.

The essential engine that has met this need for the last 40 years is now in

considerable danger, and this has serious implications for our economy,

our military, our research institutions, and our way of life.

Five decades of exponential growth in processor performance led to

It can be difﬁcult even for seasoned veterans to understand the effects of exponential

growth of the sort seen in the computer industry. On one level, industry experts, and even

consumers, display an implicit understanding in terms of their approach to application and

system development and their expectations of and demands for computing technologies.

On another level, that implicit understanding makes it easy to overlook how extraordinary

the exponential improvements in performance of the sort seen in the information technology

industry actually are.

The Future of Computing Performance: Game Over or Next Level?

54 THE FUTURE OF COMPUTING PERFORMANCE

the rise and dominance of the general-purpose personal computer. The

success of the general-purpose microcomputer, which has been due pri-

marily to economies of scale, has had a devastating effect on the develop-

ment of alternative computer and programming models. The effect can be

seen in high-end machines like supercomputers and in low-end consumer

devices, such as media processors. Even though alternative architectures

and approaches might have been technically superior for the task they

were built for, they could not easily compete in the marketplace and

were readily overtaken by the ever-improving general-purpose proces-

sors available at a relatively low cost. Hence, the personal computer has

been dubbed “the killer micro.”

Over the years, we have seen a series of revolutions in computer

architecture, starting with the main-frame, the minicomputer, and the

work station and leading to the personal computer. Today, we are on

the verge of a new generation of smart phones, which perform many

of the applications that we run on personal computers and take advan-

tage of network-accessible computing platforms (cloud computing) when

needed. With each iteration, the machines have been lower in cost per

performance and capability, and this has broadened the user base. The

economies of scale have meant that as the per-unit cost of the machine has

continued to decrease, the size of the computer industry has kept growing

because more people and companies have bought more computers. Per-

haps even more important, general-purpose single processors—which all

these generations of architectures have taken advantage of—can be pro-

grammed by using the same simple, sequential programming abstraction

at root. As a result, software investment on this model has accumulated

over the years and has led to the de facto standardization of one instruc-

tion set, the Intel x86 architecture, and to the dominance of one desktop

operating system, Microsoft Windows.

The committee believes that the slowing in the exponential growth

in computing performance, while posing great risk, may also create a

tremendous opportunity for innovation in diverse hardware and software

infrastructures that excel as measured by other characteristics, such as low

power consumption and delivery of throughput cycles. In addition, the

use of the computer has becomes so pervasive that it is now economical

to have many more varieties of computers. Thus, there are opportunities

for major changes in system architectures, such as those exempliﬁed by

the emergence of powerful distributed, embedded devices, that together

will create a truly ubiquitous and invisible computer fabric. Investment in

whole-system research is needed to lay the foundation of the computing

environment for the next generation. See Figure 2.1 for a graph showing

ﬂattening curves of performance, power, and frequency.

Traditionally, computer architects have focused on the goal of creating

The Future of Computing Performance: Game Over or Next Level?

WHAT IS COMPUTER PERFORMANCE? 55

100

1,000

10,000

100,000

1,000,000

10,000,000

1985 1990 1995 2000 2005 2010

Year of Introduction

Num Transistors (in Thousands)

Relative Performance

Clock Speed (MHz)

Power Typ (W)

NumCores/Chip

FIGURE 2.1 Transistors, frequency, power, performance, and cores over time

(1985-2010). The vertical scale is logarithmic. Data curated by Mark Horowitz with

input from Kunle Olukotun, Lance Hammond, Herb Sutter, Burton Smith, Chris

Batten, and Krste Asanoviç.

computers that perform single tasks as fast as possible. That goal is still

important. Because the uniprocessor model we have today is extremely

powerful, many performance-demanding applications can be mapped to

run on networks of processors by dividing the work up at a very coarse

granularity. Therefore, we now have great building blocks that enable us

to create a variety of high-performance systems that can be programmed

with high-level abstractions. There is a serious need for research and

education in the creation and use of high-level abstractions for parallel

systems.

However, single-task performance is no longer the only metric of

interest. The market for computers is so large that there is plenty of eco-

nomic incentive to create more specialized and hence more cost-effective

machines. Diversity is already evident. The current trend of moving com-

putation into what is now called the cloud has created great demands for

high-throughput systems. For those systems, making each transaction run

as fast as possible is not the best thing to do. It is better, for example, to

have a larger number of lower-speed processors to optimize the through-

put rate and minimize power consumption. It is similarly important to

conserve power for hand-held devices. Thus, power consumption is a